Deisy Chaves

Global and Local Features for Char Image Classification

The use of image analysis in understanding how powdered coal burns during the combustion plays a significant role in setting combustion parameters. During the pulverised coal combustion, char particles are produced by devolatising coal and represent the dominant stage in the combustion process. The pyrolysis produces different char morphologies that determine coal reactivity affecting the performance of coal combustion in power plants and the emissions of carbon dioxide, CO2. In this paper, an automatic char classification model is proposed using supervised learning. A general classification model is trained given a set of char particles classified by an expert. In particular, Support Vector Machine (SVM) and Random Forest are the trained classifiers. Two types of features are evaluated to built classification models: local and global. Local features are calculated using the Scale-Invariant Transform Feature (SIFT). Global features are defined based on the morphology classification by the International Committee for Coal and Organic Petrology (ICCP). Each classifier is trained by SVM or Random Forest and evaluated using a 10-fold cross-validation. The 70% of data is used as training set and the rest as testing set. A total of 2928 char-particle images are used for evaluating performance of classification models. Additionally, evaluation of model generalisation capability is done using a test set of 732 char particle images. Results showed that global features – defined by the application domain – increase significantly the accuracy of classifiers. Also, global features have more generalisation power than local features. Local features lack of meaning in the application domain and classifiers build with local features – such as SIFT – depend crucially on the training set.

Automatic Classification of Non-informative Frames in Colonoscopy Videos Using Texture Analysis

Colonoscopy is the most recommended test for preventing/detecting colorectal cancer. Nowadays, digital videos can be recorded during colonoscopy procedures in order to develop diagnostic support tools. Once video-frames are annotated, machine learning algorithms have been commonly used in the classification of normal-vs-abnormal frames. However, automatic analysis of colonoscopy videos becomes a challenging problem since segments of a video annotated as abnormal, such as cancer or polypos, may contain blurry, sharp and bright frames. In this paper, a method based on texture analysis, using Local Binary Patterns on the frequency domain, is presented. The method aims to automatically classify colonoscopy video frames into either informative or non-informative. The proposed method is evaluated using videos annotated by gastroenterologists for training a support vector machines classifier. Experimental evaluation shown values of accuracy over 97%.

Concave points for separating touching particles

Separation of touching objects/particles is a step before measuring morphological characteristics. An approach for identifying and splitting touching char particles is presented. The proposed approach is based on two processes. First, concave points are detected using a concavity measure and a list of touching point candidates is built. Second, separation lines are identified using location, length, blur and size. A decision criterion is derived for deciding whether or not to split a particle. The proposed approach is evaluated using 180 images of char particles and compared to the Watershed algorithm. The evaluation was twofold: quantifying the accuracy of identifying touching particles and measuring the separation quality. Expert criteria are used as a ground truth for qualitative evaluations. A good agreement between the visual judgement and automatic results was obtained, using the proposed approach.

A segmentation approach for touching char particles

Separation of touching char particles is required for measuring morphological characteristics. In this paper, a segmentation approach for touching char particles is presented. The proposed approach is fourfold. Firstly, contours are extracted. Secondly, concave points are identified by the means of measuring concavity using gradient directions at contour points. Concave points are candidates of touching point. Thirdly, separation lines are identified using location, length, blur and area. Fourthly, a decision criterion is derived for deciding whether to split a particle or not. Coal samples, from three Colombian regions (Antioquia, Cundinamarca, and Valle) and blend coals 50%-50% were devolatilised and chars were obtained. The proposed approach was evaluated using 180 images of char particles and compared to the Watershed algorithm. The evaluation was twofold: quantifying the accuracy in identifying touching particles and measuring the separation quality. An expert criterion was used, as a ground truth, for qualitative evaluations. A good agreement between the visual judgement and automatic results was obtained, using the proposed approach.